CN1279698C - Iteration terminating using quality index criteria of Turbo codes - Google Patents

Abstract

A decoder dynamically terminates iteration calculations in the decoding of a received convolutionally coded signal using quality index criteria. In a turbo decoder with two recursion processors connected in an iterative loop, at least one additional recursion processor is coupled in parallel at the inputs of at least one of the recursion processors. All of the recursion processors perform concurrent iterative calculations on the signal. The at least one additional recursion processor calculates a quality index of the signal for each iteration. A controller terminates the iterations when the measure of the quality index criteria exceeds a predetermined level.

Description

The iteration stopping criterion of the quality marker standard in the utilization Turbo code

Technical field

Present invention relates in general to communication system, specifically, relate to the decoder in a kind of receiver that is applied to the convolutional encoding communication system.

Background technology

Convolution code is applied in the digital communication system usually, is used to protect transmission information to prevent mistake.This digital communication system comprises direct sequence CDMA (DS-CDMA) standard I S-95 and global system for mobile communications (GSM).Typically, in these systems, signal is the output encoder vector that is used to transmit by convolutional encoding.At receiving terminal, there is a soft decision decoder, as the Viterbi decoder known in this area, this decoder uses lattice structure to carry out optimum search, to search out maximum likelihood transmission code vector.

Recently, the turbo sign indicating number has obtained very big development, it in addition surpassed the convolution code technology.In general, the turbo encoder comprises two or more convolution coders and turbo interleaver.Turbo decoding is adopted iteration and is used soft output (soft output) decoder to decipher single convolution code.Soft-output decoder provides information on each bit, this information can help soft-output decoder to decipher other convolution codes.Usually, the software output decoder is MAP (maximum a posteriori probability) decoder or software output Viterbi algorithm (SOVA) decoder.

Turbo coding effectively is applied in the have additive white Gaussian noise channel of (AWGN) to carry out error correction.Intuitively, there is several different methods to be used for checking and estimating the error-correcting performance of turbo decoder.Observe the carrying out along with iteration, the amplitude of the log-likelihood of each information bit of decoder iterative part (LLR) increases.This amplitude increases and has improved the probability of correct judgement.The LLR amplitude increases directly related with the iterations of turbo decode procedure.Yet, wish to reduce iterations to save computing time and to reduce the circuit power consumption.Concerning a reliable turbo decode block, the suitable number of times (stopping criterion) of iteration is along with the quality of input signal and the caused number of bit errors of transmission course is different and different.That is to say that iterations is relevant with channel conditions, the iterations that needs in the big more environment of noise is many more, to be used for correct analytical information bit and reduce mistake.

An existing stopping criterion stops decode procedure with parity check as indication.With regard to realization, parity check is direct.Yet when a large amount of bit mistake, parity check is insecure.The stopping criterion of another kind of turbo decoding iteration is that each decoding bit is calculated LLR (log-likelihood) value.Because can restrain through turbo decoding after certain iterations, and the LLR of data bit is the mark (index) that can directly indicate this convergence.A kind of method of using this stopping criterion is with the amplitude of LLR and certain threshold.Yet,, be this threshold setting that a suitable value is very difficult because channel conditions has nothing in common with each other.Existing another stopping criterion is measured the difference of entropy or two probability distribution, but this amount of calculation that needs is very big.

So, needing such decoder, this decoder can decide the suitable halt of decoder iterations in a kind of reliable mode.An advantage of this decoder is exactly can provide a stopping criterion and the complexity that can not increase computing.

Summary of the invention

According to the present invention, a kind of computational methods (100) that service quality mark standard is come termination of iterations in the convolution code signal that receives is deciphered are provided, described method comprises step:

The turbo decoder of two recursive processors with the soft input that is connected on the iterative cycles, soft output is provided, be coupling in the input of one of them recursive processor concurrently with at least one additional viterbi decoder recursive processor, all recursive processors carry out iterative computation to described signal concurrently;

For the quality of signals mark of each iterative computation in described at least one additional recursive processor;

When measuring quality status stamp above a predetermined value, termination of iterations; With

After stopping step, provide the output that from the soft output of described turbo decoder, obtains.

Brief description of drawings

Fig. 1 shows the lattice shape of the using figure of prior art in software output decoder technology;

Fig. 2 shows the simplified block diagram of turbo encoder of the prior art;

Fig. 3 shows the simplified block diagram of turbo decoder of the prior art;

Fig. 4 shows the simplified block diagram of the turbo decoder of use iteration quality marker standard of the present invention;

Fig. 5 shows the simplified block diagram of employed Viterbi decoder in Fig. 4;

Fig. 6 shows the improved figure line that hard amount of the present invention (hard guality) mark provides;

Fig. 7 shows the improved figure line that soft-quality mark of the present invention provides;

Fig. 8 shows improved another figure line provided by the present invention;

Fig. 9 shows the improved figure line that hard inside of the present invention (hard intrisic) SNR mark is provided;

Figure 10 shows the improved figure line that soft inner SNR mark of the present invention is provided; With

Figure 11 shows turbo interpretation method of the present invention.

Embodiment

The invention provides a turbo decoder, when circulating the decoding iteration, this decoder is in the input stage of each member's decoder, and dynamic utilization actual (virtual) (inside) SNR is as quality status stamp (guality index) stopping criterion of looping traffic.This quality status stamp is used as stopping criterion with the required iterations of decision decoder.Advantageously, in order to realize reliable decoding, the operand that the present invention carries out by restriction has been saved the power consumption of communication equipment and has been reduced the complexity of computing.

Typically, block encoding, convolutional encoding, coded systems such as turbo coding are illustrated as lattice shape figure as shown in Figure 1, and the lattice shape figure of Fig. 1 has shown 4 states, the lattice shape figure of 5 lattice sections (section trellis).For simplicity, we are designated as M state (M equals 8 states usually) with each lattice segment mark, are N lattice section (N=5000 usually) with each piece or frame flag.Known in this area, the decoder of maximum a posteriori type (such as log-MAP, MAP, max-log-MAP, constant-log-MAP etc.) utilize forward direction and back to Viterbi recurrence that generates or the soft output Viterbi algorithm (SOVA) on lattice shape figure, to provide soft output at every section.The MAP decoder is based on the error probability minimum of the decoding bit of all each information bits of ratio special envoy that receive.

Because coded sequence has Markov characteristic (past state can not influence state or the branch of state output in the future in the future in this sequence), the MAP bit probabilities can be divided into past state (the lattice shape of current state begins), current state (branch metric of currency) and state in future (the lattice shape of currency is finished).Especially, the MAP decoder is carried out forward direction and backward recursive, up to current state, here passes by with probability and current branch metric are adjudicated with generating output together in the future.It is known in this area that the principle of hard output and soft output judgement is provided, and exists the various distortion of above-mentioned interpretation method.Most soft inputting and soft output (SISO) decoding of Turbo code is based in the prior art by L.R.Bahi, J.Cocke, optimum MAP algorithm in the paper that F.Jelinek and J.Raviv write, this thesis topic is " OptimalDecoding of Linear Codes for Minimizing Symbol Error Rate ", be published in " IEEE Transcaction on Information Theory " Vol.IT-20, in March, 1974, pp.284-7 (bcjr algorithm).

Fig. 2 shows a typical turbo encoder, and this encoder is made of interleaver and two constituent encoders, and this constituent encoder is system's convolution coder normally, but also can be block encoder.Usually, a turbo encoder is formed by two recursive system convolution coders and the interleaver between them (int) parallel cascade.The output of Turbo coding is by with information bit m _iWith parity check bit p from two encoder RSC1 and RSC2 output _iMultiplexed (cascade) and generate.Optionally, as known in this area, can delete and cut Parity Check Bits to improve code check (1/2 throughput).Then, the turbo code signal transmits in channel.Since additive white Gaussian noise (AWGN) characteristic of channel, in transmission course, noise n _iBe added to signal X _iIn.The noise variance of additive white Gaussian noise (AWGN) can be represented as σ ²=N ₀/ 2, N here ₀The/2nd, double-side band noise energy spectral density.When receiving terminal is wanted by decoding input signal y _i(=x _i+ n _i), obtain original information bits m _iThe time, noise has increased the probability of bit mistake.Accordingly, noise effect the transmission Parity Check Bits, it is t that signal is provided _i=p _i+ n _i

Fig. 3 is a typical turbo decoder, and this decoder is made up of interleaver, deinterleaver and decoder.About external information L _E1, L _E2, interleaver (int), deinterleaver (deint) and the mechanism of carrying out the turbo decoder of iterative processing between soft input soft output decode part SISO1 and SISO2 are followed the Bah1 algorithm.Suppose in the turbo decoder to be zero decoding delay, first decoder (SISO1) is from input signal bit y _iIn calculate soft output and prior information (L _a), this prior information will be mentioned in aftermentioned.Soft output is represented as L _E1, it is as the external data (extrinsic data) from the output of first decoder.L after interweaving _E1(come from L _aPrior information) and input signal bit y _iBe imported into second decoder (SISO2) together.Second decoder has generated external data L _E2With original information bits m is provided _iThe soft output (normally MAP maximum likelihood) of soft output.External data L _E2Become L after deinterleaving _a, L _aBe fed back first decoder.Usually, for the above-mentioned iterative repetition that each bit will carry out fixed number of times (normally 16 times), all decoded up to all input bits.

The MAP algorithm can make the information bit error probability minimum of receiving sequence, and it the information bit that makes receiving sequence also is provided is 1 or 0 probability.Existing bcjr algorithm provides soft output to each bit (the lattice section of Fig. 1), and the influence of soft input is divided into soft input in the past (soft input early), current soft input and soft input (later soft input) in the future in these bits.The BCJR decoding algorithm has used behind forward direction on the lattice shape figure and one to broad sense Viterbi recurrence (generalized viterbi recursion), so that each lattice section (stage) reaches the soft output of optimum.The probability of this posterior probability or the most general log-likelihood (LLR) transmits between the SISO of iteration turbo decoding decoding step.(k=1～N), the LLR of each information bit is for all bits in the decoding sequence

${\mathrm{La}}_i=\ln \frac{\mathrm{\&Sigma;}(m,n)\&Element;{\mathrm{<figure-callout\; id="1"\; label="B"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">B</figure-callout>}}^1{\mathrm{\&alpha;}}_{k-1}\left(n\right){\mathrm{\&gamma;}}_k(n,m){\mathrm{\&beta;}}_k\left(m\right)}{\mathrm{\&Sigma;}(m,n){\&Element;\mathrm{<figure-callout\; id="0"\; label="B"\; filenames="C0180829400211.png,C0180829400232.png"\; state="\{\{state\}\}">B</figure-callout>}}^0{\mathrm{\&alpha;}}_{k-1}\left(n\right){\mathrm{\&gamma;}}_k(n,m){\mathrm{\&beta;}}_k\left(m\right)},\&CenterDot;\&CenterDot;\&CenterDot;\left(1\right)$

In equation (1), the probability that the decoding bit equals 1 (or 0) among the lattice shape figure of receiving sequence is made up of the product of some, this since yard the Markov characteristic.The Markov characteristic shows, for the current state that provides, past state and in the future state be separate.When the past state is k-1 when being n in the time, current state γ _k(n is to generate code element γ during for m for the k state in the time m) _kProbability.Current state is carried out the function of branch metric.Past state α _t(m), be at receiving sequence { y ₁..., y _kThe time be the probability of k state m in the time, state β in the future _k(k), be the receiving sequence { y that generates among time k state m _K+1..., y _NProbability.Probability α _k(m) can be expressed as α _K-1(m) and γ _k(this probability can be called as forward recursive for n, the m) function of Gou Chenging

${\mathrm{\&alpha;}}_k\left(m\right)=\underset{n=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k-1}\left(n\right){\mathrm{\&gamma;}}_k(n,m),m=0,\&CenterDot;\&CenterDot;\&CenterDot;,M-1,\&CenterDot;\&CenterDot;\&CenterDot;\left(2\right)$

M is a status number.For from β _K+1(n) and γ _k(n, the probability β that calculates in m) _k(k) contrary

Recurrence or backward recursive are:

${\mathrm{\&beta;}}_k\left(n\right)=\underset{n=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_{k-1}\left(m\right){\mathrm{\&gamma;}}_k(n,m),n=0,\&CenterDot;\&CenterDot;\&CenterDot;,M-1,\&CenterDot;\&CenterDot;\&CenterDot;\left(3\right)$

By lattice shape being schemed the B in the branch ¹(or B ⁰) addition, and obtain all posterior probability in the equation (2), B ¹(or B ⁰) corresponding with information bit 1 (or 0).

LLR in the equation (1) need can use at time k forward direction and contrary recurrence.Generally speaking, calculate and store whole contrary recurrence with the iteration of fixed number of times, and use α for realizing the BCJR method that this needs use _K-1And β _kRecursive calculation α _k(m) and L α _k(k=1 is to k=N).

The performance of Turbo decoding is subjected to the influence of several factors.One of them key factor is a number of iterations.Because the turbo decoder can restrain after the several iteration, after restraining, carry out iteration again and also can not bring great raising on the performance.Under preferable channel conditions, the convergence of Turbo code is very fast, and only needing several times, iteration just can reach good performance.Carry out number of iterations and be directly proportional, and can influence power consumption with the amount of calculation that needs.Because power consumption move and portable radio communication device in received very big concern, therefore, in turbo decoding importantly find reliable and good iteration stopping standard.Inspired by these factors, be the invention provides an adaptation scheme to be used to stop iterative process.

In the present invention, number of iterations be defined as used SISO decode stage sum (such as, iteration is twice in the circulation).Therefore, iterations from 0 to 2N-1.Each decode stage can be MAP or SOVA.The key factor of decode procedure is that external information is attached among the SISO.Then, according to the LLR value after the iteration stopping information bit is made last hard decision.Last hard bit decision is based on the polarity of LLR.For hard output, if LLR is positive, just judgement is+1, otherwise, be-1 with regard to judgement.

In the present invention, the signal to noise ratio (inner SNR) in the ring is used as the iteration stopping standard in the turbo decoder.Because when more bit in each iteration is detected as when correct, SNR will improve, the present invention has used a detection mass indicator, is used to observe the carrying out along with iteration, and signal energy is with respect to the increase of noise.

Fig. 4 shows a turbo decoder of the present invention, and this decoder has at least one additional Viterbi decoder, is used to monitor decode procedure.Though can use a Viterbi decoder, in any one SISO decoder, use two decoders can provide stopping criterion flexibly.Use the Viterbi decoder to be because can analyze the Viterbi decoder easily to obtain the quality indication.In the present invention, the Viterbi decoder only is used to do the arithmetic computing, such as, obtain quality indication and inner SNR value.Do not need to carry out real Viterbi decoding.As known in this area, if do not use iteration, MAP or SOVA will can not be better than traditional Viterbi decoder.Therefore, quality status stamp also is applied in MAP and the SOVA decoder performance and goes.Owing to Viterbi is that the error that the approximation of SISO (MAP or SOVA) produces will can not produce accumulation, this is because any change does not take place turbo decode procedure itself.Notice that turbo deciphers processing and will keep intact.Should be used to analyze with the generation quality status stamp by at least one additional Viterbi decoder, and do not need really to decipher.

In a preferred embodiment, two Viterbi decoders have been used.In practice, though when using two same decoders, thereby also need two identical SISO decoders, although can use two identical RSC, can only need a Viterbi decoder.Otherwise two Viterbi decoders are different, just need two decoders.These two decoders all can produce the iteration stopping signal, and they are separate, so any one decoder can send the iteration stopping signal.The Viterbi decoder also is not used in traditional decoding, and this is because they only are used to do the arithmetic computing and obtain quality status stamp and inner SNR value.In addition, because iteration can stop, when iteration stopping,, can from the LLR of decoder, generate soft output in the half cycles of any one SISO decoding for transmitted bit.

The present invention utilizes the external information in the iterative cycles of Viterbi decoder.For additive white Gaussian noise (AWGN) channel, with the external information input, we can obtain following path metric:

$p\left[Y\right|X]=\underset{i=0}{\overset{L-1}{\mathrm{\&Pi;}}}p[y_i\left|x_i\right]p\left[t_i\right|p_i\left]p\right[m_i]$

In the formula, m _iBe transmission information bit, x _i=m _iBe systematic bits, p _iIt is Parity Check Bits.Work as m _iWhen representing with polarity (1 →+1,0 →-1), external information can be rewritten as:

$p\left[m_i\right]=\frac{e^{z_i}}{1+e^{z_i}}=\frac{e^{z_i/2}}{e^{-z_i/2}+e^{z_i/2}},$ Work as m _i=+1 o'clock;

$p\left[m_i\right]=\frac{1}{1+e^{z_i}}=\frac{e^{-z_i/2}}{e^{-z_i/2}+e^{z_i/2}},$ Work as m _i=-1 o'clock;

P[m _i] be the prior information of transmitted bit, $z_i=\log \frac{p[m_i=+1]}{p[m_i=-1]}$ Be external information, perhaps, in general

$p\left[m_i\right]=\frac{e^{m_i{z}_i/2}}{e^{-z_i/2}+e^{z_i/2}},$

Therefore, path metric calculates like this

$p\left[Y\right|X]=\underset{i=0}{\overset{L-1}{\mathrm{\&Pi;}}}p[y_i\left|x_i\right]p\left[t_i\right|p_i\left]p\right[m_i]$

$={\left(\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\right)}^{\mathrm{<figure-callout\; id="1"\; label="L\; e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">L</figure-callout>}}{e}^{}{}^{\frac{1}{2{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\{{(x_i-y_i)}^2+{(p_i-t_i)}^2\}}\left(\underset{i=0}{\overset{L-1}{\mathrm{\&Pi;}}}\frac{1}{e^{-z_i/2}+e^{z_i/2}}\right)e^{\frac{1}{2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{m}_i{z}_i}$

Note

It is the modifying factor that external information is introduced.From the angle of Viterbi decoder, this modifying factor has been improved path metric and has therefore been improved decoding performance.External information has been brought the improvement of this factor.The present invention has introduced this factor with quality status stamp and iteration stopping standard as the turbo sign indicating number.

Concrete, quality status stamp Q (the iter, { m of turbo decoding _i, L) be:

$Q(\mathrm{iter},\{m_i\},L)=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{m}_i{z}_i$

Iter is an iterations in the formula, and L represents the bit number of each decode block, m _iBe transmission information bit, z _iIt is the external information that in each son decoding step, generates.More generally,

$Q(\mathrm{iter},\{m_i\},\{w_i\},L)=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{w_{i}m}_i{z}_i$

W in the formula _iIt is the weighting function that is used to change performance.In a preferred embodiment, w _iIt is constant 1.

Because common z _iAnd m _iHas identical polarity, so this mark is positive always.In practice, the data bit { m of input _iBe unknowable, so in fact will be with following mark:

$Q_H(\mathrm{iter},\{m_i\},L)=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{\hat{d}}_i{z}_i$

In the formula

It is the hard decision that from LLR information, extracts.Promptly ${\hat{d}}_i=\mathrm{sign}\left\{L_i\right\},$ L _iExpression LLR value.Equally also can soft output with following quality status stamp:

$Q_S(\mathrm{iter},\{m_i\},L)=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{L}_i{z}_i$

Perhaps more generally

$Q_S(\mathrm{iter},\{m_i\},\{w_i\left\}L\right)=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{w_{i}m}_i{z}_i$

Notice that producing these marks is easily, almost without any need for hardware.In addition, in fact these marks have identical progressive nature, and can be used to the better quality mark of evaluation of turbo decoding performance and iteration stopping standard.

The characteristic of these marks is that their growths are very fast in initial several iterative process, and they almost approach a constant value then.From following simulation result as can be seen, this progressive nature can well be used for describing the turbo decode procedure and be used as the quality monitoring of turbo decode procedure.In operation, if when mark value surpasses the flex point of asymptote, iteration will be terminated.

The iterative cycles of Turbo decoder has increased the amplitude of LLR, to reduce the probability of mistaken verdict.From another perspective, in fact the external information that is input to each decoder has improved the SNR that input sample flows.Following analysis will show how external information improves the actual SNR of each member's decoder.How this gain that will help to explain the turbo sign indicating number obtains.By assisting of aforementioned Viterbi decoder, also provide the analysis of input sample.

The path metric equation of attached additional Viterbi decoder is:

$p\left[Y\right|X]={\left(\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\right)}^{\mathrm{<figure-callout\; id="1"\; label="L\; e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">L</figure-callout>}}{e}^{}{}^{\frac{1}{2{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\{{(x_i-y_i)}^2+{(p_i-t_i)}^2\}}\left(\underset{i=0}{\overset{L-1}{\mathrm{\&Pi;}}}\frac{1}{e^{-z_i/2}+e^{z_i/2}}\right){\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{m}_i{z}_i}$

The expanded type of this equation is:

$p\left[Y\right|X]$

$={\left(\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\right)}^{2L}(\underset{i=0}{\overset{L-1}{\mathrm{\&Pi;}}}\frac{1}{e^{-z_i/2}+e^{z_i/2}}){\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{2{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}({{\mathrm{<figure-callout\; id="2"\; label="x\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">x</figure-callout>}}_i}^2+{y_i}^2)}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{2{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}({{\mathrm{<figure-callout\; id="2"\; label="t\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">t</figure-callout>}}_i}^2+{p_i}^2)}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{2{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}(2x_i{y}_i+2t_i{p}_i)}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{x}_i{z}_i}$

$={\left(\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\right)}^{2L}(\underset{i=0}{\overset{L-1}{\mathrm{\&Pi;}}}\frac{1}{e^{-z_i/2}+e^{z_i/2}}){\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{2{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}({{\mathrm{<figure-callout\; id="2"\; label="x\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">x</figure-callout>}}_i}^2+{y_i}^2)}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{2{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}({{\mathrm{<figure-callout\; id="2"\; label="t\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">t</figure-callout>}}_i}^2+{p_i}^2)}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C0180829400211.png"\; state="\{\{state\}\}">e</figure-callout>}}^{\frac{1}{{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}(x_i{y}_i+t_i{p}_i)+\frac{1}{2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{x}_i{z}_i}$

Observe continuous item, we obtain the following factor

$\frac{1}{{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}(x_i{y}_i+\frac{{\mathrm{\&sigma;}}^2}{2}{x}_i{z}_i)+\frac{1}{{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{t}_i{p}_i$

$=\frac{1}{{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{x}_i(y_i+\frac{{\mathrm{\&sigma;}}^2}{2}{z}_i)+\frac{1}{{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{t}_i{p}_i$

$=\frac{1}{{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\{x_i(y_i+\frac{{\mathrm{\&sigma;}}^2}{2}{z}_i)+t_i{p}_i\}$

For the Viterbi decoder, the process of the maximum correlation of search minimum Euclideam distance below search is identical.

$\frac{1}{{\mathrm{\&sigma;}}^2}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\{x_i(y_i+\frac{{\mathrm{\&sigma;}}^2}{2}{z}_i)+t_i{p}_i\}$

Suitable therewith, the input traffic of Viterbi decoder is

This diagrammatic representation such as Fig. 5.

According to standard snr computation formula

$\mathrm{SNR}=\frac{{\left(E\right[y_i\left|x_i\right])}^2}{{\mathrm{\&sigma;}}^2}$

And consider in fact y _i=x _i+ n _iAnd t _i=p _i+ n _i(p here _iBe the Parity Check Bits of input signal), the SNR that enters into the input data sample of member's decoder is:

$\mathrm{SNR}(x_i,y_i,\mathrm{iter})=\frac{{\left(E\right[y_i+\frac{{\mathrm{\&sigma;}}^2}{2}{z}_i\left|x_i\right)}^2}{{\mathrm{\&sigma;}}^2}$

$\frac{{\left(E\right[x_i{+n}_i+\frac{{\mathrm{\&sigma;}}^2}{2}{z}_i\left|x_i\right)}^2}{{\mathrm{\&sigma;}}^2}$

$=\frac{{(x_i+\frac{{\mathrm{\&sigma;}}^2}{2}{z}_i)}^2}{{\mathrm{\&sigma;}}^2}$

$=\frac{{x_i}^2}{{\mathrm{\&sigma;}}^2}+x_i{z}_i+\frac{{\mathrm{\&sigma;}}^2}{4}{{\mathrm{<figure-callout\; id="2"\; label="z\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">z</figure-callout>}}_i}^2$

Note the input owing to external information, last two is correction term.The SNR of input odd even sample value is:

$\mathrm{SNR}(p_i,t_i,\mathrm{iter})=\frac{{\left(E\right[t_i\left|p_i\right])}^2}{{\mathrm{\&sigma;}}^2}=\frac{{\left(E\right[p_i+n_i\left|p_i\right])}^2}{{\mathrm{\&sigma;}}^2}=\frac{{{\mathrm{<figure-callout\; id="2"\; label="p\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">p</figure-callout>}}_i}^2}{{\mathrm{\&sigma;}}^2}$

Along with the carrying out of iteration, the SNR of each receiving data sample value changes as can be seen, and this is because the input external information can improve reality or inner SNR.And the corresponding SNR of each odd even sample value is not influenced by iteration can yet.Clearly, if x _iWith z _iHave identical symbol, will obtain:

$\mathrm{SNR}(x_i,y_i,\mathrm{iter})=\frac{{(x_i+\frac{{\mathrm{\&sigma;}}^2}{2}{z}_i)}^2}{{\mathrm{\&sigma;}}^2}\&GreaterEqual;\frac{{{\mathrm{<figure-callout\; id="2"\; label="x\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">x</figure-callout>}}_i}^2}{{\mathrm{\&sigma;}}^2}=\mathrm{SNR}(x_i,y_i,\mathrm{iter}=0)$

This shows that external information has improved the actual SNR of the data flow that is input to each member's decoder.

Whole average SNR of each iteration phase is

$\mathrm{AverageSNR}\left(\mathrm{iter}\right)=\frac{1}{2L}\{\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\mathrm{SNR}(x_i,y_i,\mathrm{iter})+\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\mathrm{SNR}(p_i,t_i,\mathrm{iter})\}$

$=\frac{1}{2L}\{\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\frac{{{\mathrm{<figure-callout\; id="2"\; label="x\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">x</figure-callout>}}_i}^2}{{\mathrm{\&sigma;}}^2}+\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\frac{{{\mathrm{<figure-callout\; id="2"\; label="p\; i"\; filenames="C0180829400223.png,C0180829400232.png"\; state="\{\{state\}\}">p</figure-callout>}}_i}^2}{{\mathrm{\&sigma;}}^2}\}+\frac{1}{2L}\{\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{x}_i{z}_i+\frac{{\mathrm{\&sigma;}}^2}{4}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{z_i}^2\}$

$=\mathrm{AverageSNR}\left(0\right)+\frac{1}{2L}$

$Q(\mathrm{iter},\{m_i\},L)+\frac{{\mathrm{\&sigma;}}^2}{4}\left(\frac{1}{2L}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{z_i}^2\right)$

If external information has identical symbol with the data sample that receives, and if z _iThe amplitude of sample value increases, and along with the increase of iterations, whole average SNR will improve.Notice that as previously mentioned, second is the original quality mark, and second is removed by block size.The 3rd mean value with the amplitude square of external information is directly proportional, and it is always positive.This inside SNR expression formula has similar progressive nature and also can be used as the decoding mass indicator to aforesaid quality status stamp.Similar with quality status stamp, in fact inner SNR value is:

$\mathrm{Average}{\mathrm{SNR}}_H\left(\mathrm{iter}\right)=\mathrm{StarSNR}+\frac{1}{2L}{Q}_H(\mathrm{iter},\{m_i\},L)+\frac{{\mathrm{\&sigma;}}^2}{4}\left(\frac{1}{2L}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{z_i}^2\right),$

Or its corresponding soft copying as

$\mathrm{Average}{\mathrm{SNR}}_S\left(\mathrm{iter}\right)=\mathrm{StarSNR}+\frac{1}{2L}{Q}_S(\mathrm{iter},\{m_i\},L)+\frac{{\mathrm{\&sigma;}}^2}{4}\left(\frac{1}{2L}\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{z_i}^2\right)$ Initial SNR value when StartSNR represents that deciphering iteration begins in the formula.Optionally, can use weighting function here.Only last two is to monitor that the decoding quality is necessary.Simultaneously also to note having ignored the generalized constant in the inner the preceding SNR expression formula.

As review, the invention provides a decoder, this decoder can be when the convolution code signal that decoding receives, the dynamic termination of iterations computing of service quality mark standard.This decoder comprises a standard turbo decoder with two recursive processors, and these two recursive processors link to each other with iterative cycles.A new aspect of the present invention is with the input of at least one additional recursive processor Parallel coupled at one of them recursive processor.Preferably, this at least one additional recursive processor is the Viterbi decoder, and two recursive processors are soft input, soft-output decoder.Preferred, two inputs that are coupling in two recursive processors that Attached Processor is parallel respectively.All recursive processors comprise Attached Processor, and parallel carries out interative computation to signal.This at least one additional recursive processor calculates the indicia of signal quality of each iteration, and when measuring quality status stamp above a predetermined value, the indicating controller termination of iterations.

Quality status stamp is the summation of the product of external information that generates and the numerical value that extracts from LLR information in each iteration.This numerical value can be the hard decision of LLR value, also can be LLR value itself.As selection, quality status stamp is the inside signal to noise ratio of the signal that calculated in each iteration.Concrete, this inside signal to noise ratio is quality status stamp adds the quadratic sum of the external information that generates in each iteration a function.Inner signal to noise ratio can be with the quality status stamp that has as the numerical value of the hard decision value of LLR, and perhaps inner signal to noise ratio can be with the quality status stamp calculating of the numerical value with LLR value.In practice, the measurement of quality status stamp can be used the slope of the quality status stamp that is replaced by subsequent iteration.

Major advantage of the present invention is to realize on hardware easily and use flexibly.That is to say that the present invention can be used to the iteration stopping of any one SISO decoder, perhaps iteration can stop in half cycles.In addition, can decipher according to Viterbi and obtain SNR, but Viterbi decoding can not used square root functions, and square root functions can increase the complexity of circuit or similar circuit uses.On the contrary, hardware of the present invention is realized very simple.

Fig. 6 shows the simulation result of turbo decoding of the present invention.Q _H(iter, { m _i, L) and Q _s(iter, { m _i, performance L) is verified by Digital Simulation.The simulation result that shows is used for confirming the progressive nature of these marks.The performance of shown turbo decoder is based on hard label and soft mark as the iteration stopping standard.Employed sign indicating number is the CDMA2000 standard code, and known in this area, the code check of this yard is 1/3, G1=13, G2=15.With size is that 2000 frames of 640 bits move emulation, and the SNR value is respectively 0.8dB, 0.9dB and 1.0dB.As known in the art, in order to obtain actual result, the storage cutting technique of Viterbi (memory cutting technique) is passed through to learn synchronously length (synchronizationlearning length) 30 and is obtained.Hard amount mark Q _H(iter, { m _i, L) with soft-quality mark Q _s(iter, { m _i, progressive nature L) is shown in Fig. 6 and Fig. 7.

Fig. 6 and 7 illustrates the raising along with SNR, and the slope of asymptote strengthens.Positive according to expectation because high SNR can provide external information preferably in decode procedure.As can be seen, along with the raising of SNR, quality status stamp can reach progressive faster, this means that SNR is high more, and it is few more to reach the required iterations of convergence.This can find out from the analysis of Viterbi decoder.

According to the progressive nature of quality status stamp, the decoding iteration stops by percentage increase (for example, slope of a curve or derivative) that checking these marks.Below figure in the stopping criterion that uses be to be that the slope of curve of 0.03dB is judged according to threshold value.If promptly

{ Q _H(iter+1, { m _i, L)-Q _H(iter, { m _i, L) }/Q _H(iter, { m _i, L)＜0.03dB will stop iteration.Similarly, for the soft-quality mark, also can use same threshold value 0.03dB.If promptly

{ Q _S(iter+1, { m _i, L)-Q _S(iter, { m _i, L) }/Q _S(iter, { m _i, L)＜0.03dB will stop iteration.In addition, as long as these marks have surpassed a predetermined threshold value and just will stop iteration, to prevent wrong indication.Optionally, be used as at mark before the standard of iteration stopping, can utilize a certain amount of pressure iteration.

The BER performance curve of the decoder of and soft-quality mark hard based on direct use as shown in Figure 8.In this preferred example, after the application quality mark, used the minimum of 9 iteration (4.5 circulations) to force number.Having used delta threshold is the mark of 0.03dB.The maximum iteration time of using in two examples all is 16, and minimum iterations all is 9.Table 1 has been listed the mean value of iterations, to show the amount of calculation of saving.

Table 1

SNR	Use the average iterations of hard amount mark	Use the average iterations of soft-quality mark
			0.8dB	12.2910	13.2020
0.9dB	11.8265	12.7995
			1.0dB	11.4195	12.3955

According to expectation, required average iterations reduces along with the raising of SNR.In addition, because the deterioration of the caused signal integrity of iteration stopping is lower than 0.1dB.

In the present invention, inner SNR also can be used as the iteration stopping standard.Because there are confidential relation in inner SNR and quality status stamp,, shown in 10, has similar performance with the numeric results of hard inner SNR and soft inner SNR as Fig. 9.Only show in the inner SNR expression formula last two and progressive nature.

Optionally, in the ARQ system, quality status stamp and inner SNR can be used as retransmission criteria.Such as, use a lower threshold value for the frame quality, if after the iteration of pre-determined number, quality status stamp or inner SNR still are lower than this lower threshold value, will stop decoding and will send a request requiring to retransmit this frame.

Should be realized that the hardware that quality status stamp is used for the iteration stopping judgement realizes it being quite simple.Owing to have LLR and external information output at each member's decode stage, only need MAC (taking advantage of and the unit that adds up) to calculate soft mark.Equally, also only need a memory cell to come storage mark, and itself and next mark comparison are calculated to carry out slope.And the storage of all quality status stamp values of each iteration phase only needs memory cell seldom.Also need a comparing unit based on a subtracter and a divider.For hard label, before MAC, need to use slicer to make hard decision.Advantageously, according to existing design, these marks can be realized with some simple associate members.

Figure 11 shows the flow chart of termination of iterations computational methods 110 of the present invention, during this calculating is used to decipher and service quality mark standard the convolution code signal that receives is deciphered.The first step 102 provides has two turbo decoders that are connected in the recursive processor on the iterative cycles, and at least one additional recursive processor Parallel coupled is at the input of one of them recursive processor.What all recursive processors were parallel carries out iterative processing to signal.In a preferred embodiment, at least one additional recursive processor is the Viterbi decoder, and two recursive processors are soft input soft output decode devices.Preferred, two inputs that are coupling in two recursive processors that Attached Processor is parallel respectively.

Next step 104 calculates the quality of signals mark for each iteration at least one recursive processor.Concrete, quality status stamp be the external information that from recursive processor, generates with the product of the LLR value of information that in each iteration, from recursive processor, extracts and.Quality status stamp can be a hard value or a soft value, and for hard value, this value is the hard decision of LLR value.For soft value, this value is a LLR value itself.Optionally, quality status stamp is the inside signal to noise ratio (snr) of the signal that calculated in each iteration.This inside SNR is that initial signal to noise ratio adds quality status stamp, adds the function of the quadratic sum of the external information that generates in each iteration.Yet for quality marker standard, it is useful having only back two.For this reason, also have hard value and soft value to inner SNR, it uses the hard decision and the soft-decision of corresponding aforementioned quality status stamp.

Next step 106 is when measuring quality status stamp when having surpassed predetermined value, termination of iterations.Preferably, stop step and be included in the quality status stamp that measurement is represented with the slope of quality status stamp on the iteration.In practice, predetermined value is on the flex point of the quality status stamp of close asymptote.More specifically, this predetermined value can be set to the 0.03dB of SNR.Next step 108 output that provides the soft output by the turbo decoder to obtain, this output are present in and stop after the step.

Though foregoing description the special composition and the function of turbo decoder that convolution code is deciphered, those skilled in the art can use its still less or additional function in the scope of broad of the present invention.The present invention is only limited by the claim of being added.

Claims (9)

1. computational methods (100) that service quality mark standard is come termination of iterations in the convolution code signal that receives is deciphered, described method comprises step:

The turbo decoder of two recursive processors with the soft input that is connected on the iterative cycles, soft output is provided, be coupling in the input of one of them recursive processor concurrently with at least one additional viterbi decoder recursive processor, all recursive processors carry out iterative computation (102) to described signal concurrently;

For the quality of signals mark (104) of each iterative computation in described at least one additional recursive processor;

When measuring quality status stamp above a predetermined value, termination of iterations (106); With

After stopping step, provide the output that from the soft output of described turbo decoder, obtains (108).

2. according to the method for claim 1, described first provides step to comprise: two Attached Processors, these two Attached Processors are coupling in the input of described two recursive processors respectively concurrently.

3. according to the method for claim 1, described calculation procedure comprises: described quality status stamp be the external information that in each iteration, generates with the product of the numerical value that from log-likelihood information, extracts and.

4. according to the method for claim 3, described calculation procedure comprises: described numerical value is the hard decision of described log-likelihood value.

5. according to the method for claim 3, described calculation procedure comprises: described numerical value is described log-likelihood value.

6. according to the method for claim 3, described calculation procedure comprises: described quality status stamp is the inside signal to noise ratio of the signal that calculates in each iteration, and described inner signal to noise ratio is the function that described quality status stamp adds the quadratic sum of the external information that generates in each iteration.

7. according to the method for claim 6, described calculation procedure comprises: the quality status stamp that described inner signal to noise ratio usefulness has as the hard decision value of described log-likelihood value calculates.

8. according to the method for claim 6, described calculation procedure comprises: the quality status stamp that described inner signal to noise ratio usefulness has as the numerical value of log-likelihood value calculates.

9. according to the method for claim 1, described termination step comprises: the quality status stamp of described measurement is the slope of the quality status stamp on iterative process.

Images